The increasing availability of online audio/visual academic lecture material enables new and exciting ways for disseminating knowledge that can potentially change the way people learn. Ideally automatic methods can be developed to enable fast, accurate and easy access to lecture content. Towards this goal, the MIT Spoken Lecture Processing Project aims to develop new techniques in the areas of automatic speech recognition, structure induction, and summarization that will enable efficient search, retrieval and browsing of audio and/or audio/visual lecture materials. The technical language of academic lectures and lack of in-domain spoken data for training makes automatic speech recognition of lectures a significant challenge.
Furthermore, the processing of potentially errorful transcripts from spontaneously generated lecture speech can significantly affect the performance of standard indexing, segmentation and summarization techniques. In this talk, we will provide an update on our current research activities and progress in these areas.